AttMoE: Attention with Mixture of Experts for remaining useful life prediction of lithium-ion batteries

Abstract

For Lithium-ion (Li-ion) batteries, problems such as material aging and capacity decay lead to battery performance degradation or even catastrophic events. Predicting Remaining Useful Life (RUL) is an effective way to indicate the health of Li-ion batteries, which helps to improve the reliability and safety of battery-powered systems. We propose a novel neural network, AttMoE, which combines an attention mechanism with Mixture of Experts (MoE), to capture the capacity fade trend for battery RUL prediction. When facing the problem that raw data collected from sensors are always full of noise, AttMoE uses a dropout mask to denoise the raw data. For RUL prediction, one key idea is that the attention mechanism captures the long-term dependencies between elements in a sequence and more attention is paid to the important features that contain more degradation information; another key idea is that MoE uses many experts to increase model capacity to achieve better representations. Finally, we conducted experiments using two public data sets to show that AttMoE is effective in RUL prediction and achieves up to 10%–20% improvement in terms of Relative Error (RE). Our projects are all open source and are available at https://github.com/XiuzeZhou/RUL.